Repository navigation
Actor servers take thousands of clients on one thread and replay only unfinished messages, and module functions keep their locals apart once they run hot
- A function from an imported module no longer mixes two of its locals from its second call on, a phi in the lowered code read its sources at the wrong instruction since 0.9.0.
- A consistency suite holds 500 generated programs to one answer per call, cold, warm, from a module and as a method, and a std suite runs the edge-python-std tests on every push.
- An eval waiting on a sleep or a host call parks instead of holding the scheduler, a busy actor yields every preempt back-edges and sits in the ready queue once, and limits.timeout sets the deadline of an eval group up to 300 seconds.
- One mio thread serves the TCP ingress and the HTTP control port, a message holds at most 16 MiB, a request has 30 seconds to arrive, and a full inbox stops reading and answers 503.
- The durable log marks each message done, compacts past 64 MiB and syncs before a /pub answers 202, so a restart replays only what was left, and SIGTERM gives the actors 25 seconds to finish.
- Dead code and the limits.heap and limits.calls notices are gone, the Lucide fetch falls back to a second CDN, and the crates say 0.9.6.